The Clinician Scientist

I'll start where I am now and then work back. I'm a clinician, and I'm also a researcher in artificial intelligence, and I currently work for Google Japan. But I've taken a few interesting turns to get here.

I originally trained as a doctor, and I thought I was going to be an academic radiologist. I got involved in research during medical school, mostly in computational analysis of MRI, and once I graduated I carried on this project — a bunch of work with a Japanese hospital at the time. And it became really obvious that when you went to conferences, the same thing was presented year in and year out. The same paragraphs of the discussion in different research papers were just lifted from paper to paper: "this could be used for this, and future work should do this." And it's like, well, is anyone really going to do it?

Around the same time, deep learning was becoming a really big thing in medicine — this was 2013-ish — and it seemed like the work I was doing in computational analysis was going to be almost completely superseded by it. So when I was picking what to do next, whether to do a PhD, deep learning was the natural thing to go into. I did a bit of reading on my own, did a few courses, and had every intention of doing a PhD at UCL.

And at the very beginning of this, even before I signed on, I was introduced to DeepMind — initially as a way of funding the research work I was doing at UCL. DeepMind was this exciting company setting up a team to apply artificial intelligence to healthcare: imaging, health records, a bunch of other stuff. It seemed too good an opportunity to pass up. I thought, well, I could do the PhD, and in four years maybe this job will still be here, maybe it won't. Or I could just skip that whole process and go and do this now.

So I learned a lot on the job and had a fantastic time, working with some amazing people. During my time at DeepMind I helped build up a whole bunch of different projects — nominally a clinician scientist role, but working across everything from coding to business development. Then I moved across when the DeepMind science team was set up. That was a very cool job, because the mandate was essentially: go and find something that will win a Nobel Prize. Go and start a bunch of different projects. Those are things I can't talk about so much here, because the publications are still ongoing, but it was looking more at basic science with a healthcare outlook.

And my current role at Google Japan combines all of these aspects of my career so far. We're looking at how we can move a lot of this research into deployments, and where there are basic science projects that might be interesting — all with a view of how we fundamentally improve and change the way healthcare works. So I'm super excited about the next chapter, and also very excited about a lot of the work I've done so far.

There's a high-level "here is what we want to achieve," and that's always going to be about impact — ideally impact in the real world. Exciting research that, at least from the work I was doing, has a real-world application that could genuinely change things. For healthcare that was very focused around patient impact. So a lot of it was: we could do a thousand different things, but which is the one that would really address a particular unmet need? Which is the one where, if we put a big team on it, the end result is something genuinely useful that can positively impact people's lives?

That was the same basic principle whether it was a project focused purely on healthcare, or something more at a basic science level, or even a more fundamental methods paper focused on a particular architecture — it still needed to link to something in the real world. There was a lot of discussion and a lot of support to take an idea that might have been half-baked to a stage where it really could become a project. So I wouldn't say it's a free-for-all, but there's definitely a lot of freedom at the idea stage, and then a lot of support to take that idea to something more impactful.

I've always been interested in reading outside of what I'm supposed to be reading. You gave me this warning before that one of the questions would be what book I recommend — and the books that have influenced me most are the ones I just probably shouldn't have been reading. I'd get books out on economics, or journalism, or physics, and maybe I wouldn't understand 90% of the book, but whatever — it was some additional thing to change how I then viewed the research questions I was confronted with in medicine.

It was the same process that got me into research in the first place. I intercalated in medical physics, and the reason I did was because I was really crap at physics. It was the worst subject when I did my GCSEs; I dropped it for A-levels, thinking I don't want to do physics ever again. And then I got this opportunity to intercalate, and it was like, well, what can I not do? What's going to be the thing I can actually learn from, because I don't want to do something I already know? So physics was the thing I chose. And that ended up being really interesting, because I knew nothing — I was a complete idiot again — but that also meant I could ask lots of interesting questions, and eventually I got involved in the computational analysis research.

We had one paper, and I wrote to a bunch of hospitals: can I apply this with you, do you have a data set? Again, before the days of machine learning, but it's really the same principle each time. When I'm looking at what to do, I enjoy learning new things, and I try to pick the most random thing that's furthest away from what I can currently do.

The group of papers we're talking about — the ones I worked on at DeepMind — are almost exactly this type of cross-pollination. They're the intersection of clinical medicine and deep learning. And I found it incredibly useful to be a person who knows a very small amount about both, working with teams who know an awful lot about one of them specifically.

They are fantastic — both jobs have been absolutely fantastic. DeepMind is a really incredible company, and with all of this it's the focus on people. Having the opportunity to spend an awful lot of time thinking about what the future should look like and how we get there is a huge privilege. And equally, both organisations focus on the individual person and making sure you're in a state where you can do that kind of research. So I'm very privileged there.

How do you get there? There's this great quote from the Simpsons — the two greatest words in the English language: "de-fault." I was just in the right place at the right time. I happened to be a person who had a little bit of knowledge in both areas. But it goes back to the point I mentioned: I don't think it's helpful to pick out "here's this great company, how exactly can I get there." It's more, what do you want to do? What's interesting to you? Go really deep into that, and then frame the problems and questions you're faced with within the context of the knowledge you've assimilated. That won't always be successful, but it's about catching the wave. At some point, if you're living your life with this mindset, a wave will come along, and you'll be prepared to catch it, because you've followed this mantra.

Yeah, along with an awful lot of other people. The field is moving as a whole, with a lot of companies, and I'm happy to have been able to contribute a little bit — but there's a huge number of people doing incredible work. One of the benefits I see from Google in particular is that, alongside developing our own models and products for healthcare, we can also enable many others through funding, or through open-source software. As I moved into Google and learned more about the business, I became quite proud to be part of this wider piece — it isn't just about being the one person riding the wave doing something cool, it's about enabling everyone to ride it.

At the moment it's a bit different because of COVID. My average day starts at 2am, because I work on Japan time for the first half of the day. I'll have meetings between about 2 and 7, take a break, have a few more meetings at night, and then I have the rest of the day to really do what I want. It's a little bit hardcore at the moment, but it does mean I have a huge amount of time to read about different areas.

Recently I've been reading a lot about healthcare — in my case, how the Japanese healthcare system works — but in a lot more depth than I've had a chance to over the last three or four years. So I can look not just at specific areas in medical imaging but really understand how different sections of a healthcare system work together. How does public health really work? Who sets the policy? How do the financial incentives work in Japan versus the US versus the UK, and how does that relate to the unmet needs? That's really interesting, because you can uncover that the unmet needs — how you can best support patients — are completely different between, say, Japan and the US.

That would be my afternoon. But next week I might spend time reviewing papers, going deep into more specific AI-applied-to-healthcare work, or diving into a completely different area again. So it's very varied — that's my currently weird COVID day.

That's a super interesting question. I think "false" is a very strong term — you have to go beyond the headline. An awful lot of the papers in deep learning, specifically in healthcare, will be hyped up more in the media and with headlines than maybe they should be. But when you read how people are discussing them and engaging at conferences, it's the sort of incremental progress towards a shared goal that's how research is supposed to work. Within any field you'll get good papers and bad papers, and often papers where there are bits that are good and bits that are bad. That's just part of research.

There are definitely areas we could improve. The field hasn't focused as much on bias and diversity as it should have. There's a lot more consideration we need to give to AI safety, and to understanding how people engage with models — how we make sure clinicians are making the right decision based on the information, and that we're not displaying an output in a way that might bias them. I wouldn't necessarily say these are failures of existing papers, but they're areas where more research is needed.

A year and a half, two years ago, I would have said basic sciences — and I think that's still kind of true. The work on trying to identify novel biomarkers, on how we better use metabolomics, how we better use emerging fields like microbiome information, and to a lesser degree drug discovery. The field for these types of questions is where image analysis was maybe five years ago.

It's not medical imaging specifically, but if I were going into something very specific now, I'd probably look at metabolomics or the microbiome. A lot of progress could be made there, particularly where you have multimodal data sets. One example — I don't know if anyone is working on it, but it would be super interesting — is trying to understand the impact of the microbiome on a person's physiology, and whether there's a way to produce a data set that lets machine learning really do anything there, maybe through biobanks. That's one of many ideas.

It's a bit of both, but the focus is more on what can help the most people. This is summed up in a lovely quote: fall in love with the problem, not the solution. The main focus is where the unmet needs are, and how you can improve the quality of healthcare in a way where it's really the right place for someone like Google to contribute. That said, I've definitely been involved in papers focusing more on fundamental methods and understanding their application to healthcare. So there probably is a need to look at both.

Review articles are super helpful here. One of my colleagues at DeepMind wrote a review for Nature Genetics on deep learning in genomics, and although it's tailored to genetic information, the explanation of how the machine learning techniques work — the diagrams in there — make it a really great article for understanding some of the basics. There are a lot of surveys on deep learning applied to medical imaging, and on deep learning applied to EHR, just to understand the type of questions people are asking.

After you've read a few of these, you can start to parse a bit more of the information. Then it's a case of realising you won't understand everything in a paper — there are very few papers where I really understand everything. You read and reread, and you understand the basic concepts. It's also useful to realise that not everything is super well written, because most of the people who write them are not professional writers, they're researchers. It does take time to learn the type of things people consider in a particular field. But it's a case of perseverance, and reading review articles to get a basic grounding.

The classic one is that improving your model will improve your performance — or that more data will. That better models or more data are really going to make the difference. Focusing on the quality of a particular data set is much more important than just adding lots and lots more, or spending lots of time tuning a model.

With all our projects, we spent a lot of time really understanding a particular data set, and making sure the information we included was as reliable as possible. The great example — and I'm sure Pearse, who you had on your podcast a while back, might have mentioned this — is that one collaboration started with something like 1.3 million OCT scans. By the time we'd found the set that matched the particular use case we cared about, filtered down to only those with sufficient quality and reliable enough labels, it was down to something like 15,000 images. That was a long process of figuring out exactly what we wanted to do, but it's a really stark example of why data set quantity is not as important as data set quality.

That's a really easy answer. The people I was able to work with at DeepMind were, in my mind, the right people to be working with. Everyone is fantastically talented. If you were to try to hire this type of team for a startup, you'd spend years and years just looking for these people, and you'd never do any work. For me it was about how I grow myself, and how I make sure I'm going to do something impactful that works — and that's being around this particular group of people. Throughout my career, the most important thing has been making sure you're working with a really good team. I found that at DeepMind, and now at Google as well.

One of the books I found really useful recently was Introduction to Genomics by Arthur Lesk. It's quite deep into genomics, but I found it super insightful — it's an area that's going to be more and more important for clinical practice, and something I felt I lacked. That depth of knowledge was definitely missing from my medical school education. And completely separately, I very much enjoyed Sapiens, which is more of a cliché now because it's so popular. But this changes on a daily basis — I'm sure I'll read something next week and say, oh no, you should read this instead.